Why does this type instability lead to allocations?

[user664303]

Many thanks for your input, @cjdoris.

@cjdoris

My guess is that, even though the return type of mysum(v) is inferred (as Float64), it is still boxing the return value, and therefore still allocating. In principle, Julia could indeed compile versions of these methods to return Float64 natively, but I guess that in the face of dynamic dispatch, Julia currently always boxes the result. Presumably this is an optimization that could be implemented.

Interesting. I felt confident that specifying the output of mysumv() would remove this boxing. If this is the source of the allocations, then as you say, it’s likely something that can be fixed. Can you not tell where boxing is occurring in the output of @code_warntype? I assumed it was on red lines of code, but the only red line in the for loop is at %7, which doesn’t relate to the output of mysumv().

@cjdoris

To summarise, it appears to take ~10ns to box the intermediate results, ~60ns to do dynamic dispatch and ~20ns to do the actual work. Hence the allocations coming from boxing have a fairly negligible impact on run-time.

If this is the case then I feel that the time taken for dynamic dispatch is something that can be significantly improved also, based on my previous tests.